Evidence · Local synthetic checks
Inspect the claim.
Reproduce the check.
On October 26, 2026, the existing mock policy’s local Node.js test suite passed eight tests, with zero skipped or TODO tests. This is evidence about the supplied evaluator and fixtures, not a production deployment or an effectiveness benchmark.
Synthetic policy only · No real actions · No model calls
The local run was observed at , using Node.js v25.9.0. View the fixture manifest and observed decisions.
What was checked
- Public synthetic reads versus lexical path bypass and protected paths.
- Exact example recipient domain and defined fake sensitive markers.
- Unknown tools, extra fields, forged approval contexts, and malformed records.
- One export CSV deletion scope and separate simulated approval.
- Exact-request approval binding, one-time use, and action separation.
- Approval cannot override a denial or survive mutation of captured arguments.
- Deterministic scenario decisions; a synthetic injection note grants no authority.
- Fingerprint argument fidelity and object key-order independence.
Reproduce locally
Download the three adjacent files below into one local directory. With a compatible recent Node.js runtime, run this command there. It needs no package installation, API key, or production access.
node --test policy.test.mjsSource identity
These SHA-256 hashes identify the exact source files linked in this edition. Hashes identify bytes; they are not a security certification.
policy.mjs04299a5b9790e90797a9e2d6f9fbfdb1327b5ee22d86d579ebc2241a9410ca65
scenarios.mjs671f5fbb0ee5ae87b93b2275ebc5d50a624bfdfdbe03f1b391517075cbf0e5d6
policy.test.mjs125a463c6ce6c50f1e45883a5a204517d6aa18801a50efe639c83418dfe7f8db
What this does not establish
The tests do not run a real tool, read or delete a file, send an external message, spend money, make a network request, or call a model. They do not establish resistance to every injection, secret disclosure, encoding, race, symlink, tenant isolation failure, or alternate execution path. A browser-side simulator can be modified by its user.
The demonstration uses lexical path rules and known fake markers. Its indirect-injection fixture is denied by recipient policy, not by understanding the malicious document. Request fingerprints are canonical serialization, not cryptographic approval credentials.
Product status and open work
Boundary is a prelaunch commercial model router. Planned request-path checks are separate from this mock agent-action demonstration. Trusted server-side enforcement, authenticated approval, atomic redemption, real resource resolution, integration coverage, and operational validation remain requirements for any real action-control implementation. Read the precise product scope.
Try one synthetic task